Predicting the Future of IT Infrastructure: Zabbix and AI Model

Predict the future, identify anomalies. It sounds like the tagline of a sci-fi movie, doesn't it? I could have used some flashy special effects to introduce this topic, but we are scientists, not science fiction writers.

Today, we are diving deep into a highly specific and fascinating topic: applying Artificial Intelligence to my absolute favorite field—Zabbix. We are exploring the cutting edge of observability and infrastructure monitoring. A question I hear almost constantly is: "Okay Dimitri, we can set thresholds and monitor everything, but I don't even know what I should be monitoring to catch a problem before it happens. How do I do that?"

It’s a fantastic question. What tools do we actually have to predict the future of our metrics? Let’s break it down.

The Legacy Way: Zabbix Forecasting and Baselines

Can Zabbix analyze and predict metrics out of the box? Yes, absolutely. Ever since version 6.0, Zabbix has included features for baseline monitoring, anomaly detection, and forecasting. We’ve successfully activated these features for many clients, especially for rigid, predictable environments like monitoring file system usage.

However, if you look at the official documentation and the whitepapers behind these features, you quickly realize the fundamental problem: it is highly technical.

To use traditional forecasting, you need to:

It’s not easy. Even for seasoned professionals, finding the perfect configuration requires a lot of trial and error. Furthermore, traditional methods only analyze a single metric in isolation, completely ignoring how different applications and environmental variables might influence each other.

The Game Changer: The TOTO Model

A few months ago, I discovered something incredibly exciting. A new, open-source AI model was released that functions similarly to Large Language Models (LLMs), but it is specifically trained on mathematical and statistical time-series data.

It’s called TOTO, which stands for Time Series Optimized Transformer for Observability. We have Datadog to thank for this kind gift. They took their massive database of infrastructure metrics, threw a few million dollars at the problem, and released a beautiful open-source model for the rest of us mere mortals.

Why TOTO Changes the Paradigm of Observability

Predicting the future of IT infrastructure is difficult because there are simply too many variables in play. TOTO tackles this head-on with several state-of-the-art features:

Putting Theory into Practice: Zabbix AI Prediction

Theory is beautiful, but I am an engineer—I need to see it work in the real world. So, I built a custom application called Zabbix AI Prediction.

Here is how it works: Let’s say you want to identify performance problems on a specific database host, but you have no idea which of the hundreds of available metrics you should actually monitor. Through my interface, you simply select the host and type a natural language prompt asking the AI to identify database performance bottlenecks.

The AI analyzes all the metrics inside that host and suggests the most critical ones to use for prediction. Once you hit "Deploy Bundle," the application hooks directly into Zabbix, tags the relevant items, and creates specific items that will receive the forecasted data.

Real-World Results

I ran a simulation using a Docker container machine to test the 60-minute forecast. By plotting the real-time data alongside the TOTO model's historical predictions and future forecasts, the results were stunning.

The model accurately predicted a massive spike (a climb ramp approaching 100% utilization) a full hour before it actually happened. Because TOTO correlates all the metrics simultaneously, we aren't just guessing; we are getting a highly accurate, multivariate prediction of system volatility.

This allows us to create highly specific, predictive triggers in Zabbix. Instead of reacting to a database crash, you get an alert 60 minutes in advance telling you exactly which correlated metrics are about to cause a bottleneck.

Conclusion

In my opinion, the introduction of models like TOTO represents a massive paradigm shift for Zabbix and the entire infrastructure monitoring industry. What was once almost impossible for everyday system administrators—true, multivariate anomaly prediction—is now highly accessible.

We finally have a solid, intelligent foundation to start predicting anomalies within our infrastructure before they impact our users.

But I want to hear from you! What do you think about integrating AI transformers into your monitoring stack? Let me know in the comments below if you find this useful, and don't forget to leave a like if you enjoyed the breakdown. I'll see you next week with a new video!

Stay Connected:

— Dimitri Bellini